Embedding providers

Hub: documentation home. Field reference: tools.yaml.

defaults.embedding may be a string (fastembed/BAAI/bge-small-en-v1.5) or an object. Per-tool query.embedding wins. ${VAR} is allowed only under config.

Provider	Extra	Config
fastembed	store extra / embed	model
openai	vectorsmith[embed-openai]	api_key, base_url, timeout_s, batch_size
azure_openai	vectorsmith[embed-openai]	api_key, endpoint, deployment, api_version
http	(httpx, built-in)	base_url, api_key, headers, request, response
cohere	vectorsmith[embed-cohere]	api_key, model

defaults:
  embedding:
    provider: openai
    model: text-embedding-3-large
    dims: 3072
    config:
      api_key: ${OPENAI_API_KEY}
      timeout_s: 30

HTTP gateway

defaults:
  embedding:
    provider: http
    model: my-model
    dims: 1536
    config:
      base_url: ${EMBED_BASE_URL}
      api_key: ${EMBED_API_KEY:-}
      timeout_s: 120
      batch_size: 32
      request:
        model_field: model
        input_field: input
      response:
        vectors_path: data[*].embedding

Default JSON body is {model, input: [texts]}. Default vector path is OpenAI-shaped data[*].embedding. dims is required for unknown model ids (VB2018 without it on --live).

VB2019 if the provider extra is missing. validate --live-embed smoke-tests the embedder. HTTP GET /readyz checks provider health on every loaded project that has search/pipeline tools, or when you pass --live-embed.

Query expansion

Separate from embeddings. query.expand (openai / http / none) rewrites the user string, then this embedder runs on each variant. Off by default. Failure → original query + VB4030. YAML: tools.yaml → query.expand.